France's telecoms and digital regulator, ARCEP, has opened a public consultation on a draft decision that would require generative AI service providers to report the greenhouse gas emissions associated with their services, the main characteristics of their most-used models in France, and the compute volume, cumulative processor time, and energy consumption tied to both training and inference. The consultation, announced July 24, 2026, runs through September 30, 2026, with a final decision expected by year-end, subject to sign-off from the minister responsible for AI and digital affairs.
This is not a new regulatory apparatus built from scratch. ARCEP has run its annual "Pour un numérique soutenable" survey since 2020, collecting environmental data first from telecom operators, then terminal and network equipment makers, then — after the 2021 REEN law and the 2024 SREN law — cloud computing providers. The 2026 edition of that survey found data center electricity consumption in France rose 38% over three years, even as efficiency per unit of compute improved. Folding generative AI providers into the same reporting structure is the next incremental step, not a regulatory leap.
The case for the extension
The strongest argument for ARCEP's move is one it made itself in a May 2026 report on generative AI's environmental footprint, produced with PEReN (the Pôle d'expertise de la régulation numérique): there is currently no reliable, comparable data on how much energy AI training and inference actually consume. The International Energy Agency projects global data center electricity demand could double between 2024 and 2030, and ARCEP's own researchers found that the largest models are consistently the most energy-intensive — while some smaller models deliver comparable output at a fraction of the energy cost. That finding matters for the innovation debate: it suggests that emissions reporting, done well, need not force a tradeoff between capability and efficiency. It can instead surface which providers are already doing the optimization work quietly, and let developers, regulators, and enterprise buyers reward it. A regulator that has no data cannot design good policy, subsidize the right things, or resist worse interventions dressed up as climate action.
Where the draft is still thin
That argument holds up better in principle than the draft decision does in its current, still-unsettled form. ARCEP has not published a revenue or usage threshold for which generative AI providers are actually in scope — contrast this with the parallel cloud-provider collection decision, which set clear €10 million French-revenue and 100 kW infrastructure-power floors. Nor does the current text specify any consequence for a provider that simply declines to respond; ARCEP's own language describes a "collaborative approach" rather than an enforcement mechanism. For a consultation that touches training-compute and inference-energy figures — the closest thing an AI lab has to a trade secret about model efficiency and scale — that ambiguity cuts both ways. Vague scope invites either uneven compliance (large labs volunteering polished numbers while smaller or foreign providers ignore the request entirely) or, if ARCEP tightens the rule later without fresh consultation, obligations that landed on firms who had no reason to expect them.
A second reporting regime, not the first
The timing also matters because ARCEP is not writing on a blank page. Under the EU AI Act, providers of general-purpose AI models already have to draw up and maintain technical documentation covering the model's known or estimated energy consumption, a duty that applies now to models placed on the market after August 2025 and phases in for earlier ones. If ARCEP's eventual French collection asks for materially different metrics, in a different format, on a different timeline, it adds a second compliance track for the same underlying disclosure rather than reusing the first. The efficient version of this policy would map ARCEP's indicators onto the AI Act's Annex XI documentation categories wherever they overlap, so a lab filing once satisfies both regimes. A ChannelNews report on the draft on the consultation notes ARCEP has been in direct exchange with generative AI providers since early 2026 to define workable indicators — a sign the agency is aware of this risk, even if the published draft doesn't yet resolve it.
The right instrument, used carefully
On balance, extending an existing, well-understood survey to a new class of provider is the proportionate move, not an overreach. ARCEP is not proposing caps on training runs, mandatory efficiency targets, or French-specific model audits — it is asking for numbers it can publish, the same posture it has taken with telecoms operators and cloud providers for five years running. That is meaningfully different from the kind of AI-specific licensing or pre-approval regimes some jurisdictions have floated. The open questions — scope thresholds, non-response consequences, and alignment with the AI Act's own energy-disclosure duty — are exactly the kind of detail a consultation period exists to fix. Whether this becomes a low-friction data point or a duplicate compliance burden will be decided in the next two months, before the final text goes to the minister at year-end.